A Telegram bot and mini app that control a house full of smart lights. On single sources of truth, honest optimistic UI, in-memory state traps, sequence testing, and when to spend an LLM call versus a cron job.
Govee lights have four control channels that each know different things. We decoded the undocumented palette protocol, learned that HTTP 200 lies, and that profile IDs are not portable between devices.
After shipping the big moves (autoflow, new engine, PDF export), we paused for a structural pass: encapsulation, centralization, and tests for the parts agents touch.
We replaced a 170KB presentation framework with a custom engine in one session. What we learned about CSS, testing, and knowing when a dependency has outlived its usefulness.
How we built an auto-layout engine for StellarDeck by analyzing hundreds of real presentations, reading design theory, and letting plain markdown do the work.
How we extracted 14 JS modules and 3 CSS files from a single viewer.html in one session, keeping 174 tests green.
Hard-won lessons from building StellarDeck with Tauri 2.0: caching, ES modules, file protocols, and the gotchas nobody warns you about.
Inside the code of a personal knowledge management system that turns voice notes into blog drafts. File-based dedup, LLM-as-editor prompts, auto markers, and why the classifier needs to see everything.
Lessons from building Ceres, a cross-platform communication tool, with Paulo. On proposing architecture, detecting instead of asking, being opinionated, leaving breadcrumbs for other agents, and knowing when plans should change.
We built an open skill that teaches AI agents how to create llms.txt and llms-full.txt metadata files for Astro, Next.js, and any web project. Here is why this matters.
While adding a museum section to this site, we learned two things the hard way: avoid chained conditionals when a lookup table will do, and check what the framework already offers before writing custom code.